The Face Recognition Algorithms Based on
Weighted LTP
Haifeng Zhang and Shenjie Xu Hangzhou Dianzi University, Hangzhou, China
Email: [email protected], [email protected]
Abstract—Local Ternary Pattern (LTP) is usually applied
for texture classification problems. LTP extends the Local
Binary Pattern using the custom threshold and encoding the
small pixel difference into third state. Since the amount of
information in different face regions is not equal, this paper
proposes an approach of weighted LTP to show facial
feature effectively. First, the original face image is divided
into small blocks, and the LTP characteristic value and
histogram of each piece of pixel are calculated. Then the
weight of sub histogram is calculated by information
entropy and the histogram of whole face image cascade of
the histogram of all sub regions, finally, the weighted
histogram of whole face image similarity are calculated by
chi-square distance, the classification is performed by a
nearest neighbor classifier. Experimental results show a
better performance on ORL and Yale face database.
Index Terms—local ternary pattern, local binary pattern,
face recognition, histogram
I. INTRODUCTION
Face recognition has received a great deal of attention
from the scientific and industrial communities over the
past several decades owing to its wide range of
applications in information security, access control and
law enforcement. The key to face recognition algorithm is
extracting the facial features. LBP as a method of
describing textures was applied to face recognition for the
first time by Ahonen et al. But the original LBP method
thresholds all pixels in the neighborhood based on the
gray value of the central pixel [1]. As a result the original
LBP becomes more sensitive to noise especially in near
uniform or flat areas. In order to solve this problem, LTP
that extract the features based on 3-valued texture
operator was proposed by Tan et al. [2].
Both LBP and LTP method divided the facial image
into some sub regions, but these regions has different
amount of characteristic information, for example, eyes
as an important characteristic of face contain more
characteristic information than others regions of face.
Therefore, that sub regions of face were given different
weights can better performance characteristic of face. In
this paper, we proposed weighted LTP algorithm that
information entropy was introduced to calculate the
weight of sub regions of face.
Manuscript received June 19, 2015; revised October 30, 2015.
II. LOCAL BINARY PATTERN
The original LBP operator, introduced by Ojala et al.,
was designed for texture description [3]. The operator
labels the pixels of an image by thresholding the 3*3
neighborhood of each pixel with the center value and
considering the result as binary number. Then, the
histogram of labels can be used as a texture descriptor [1].
Formally, the LBP operator takes the (1):
7
0
( , ) 2 ( )n
c c n c
n
LBP x y S p p
(1)
where in this case n runs over the 8 neighbors of the
central pixel c, Pn and Pc are the gray-level values at c
and n, and S(u) is 1 if u 0 and 0 otherwise. See Fig. 1
for an illustration of the basic LBP operator.
Figure 1. LBP operator.
Later the LBP operator was extended to use different
size of neighborhood to deal with different scales of
textures. Defining the neighborhood as a set of sampling
points on circle centered at the pixel to be labeled allows
any radius and number of sampling points [4]. However,
after rotation, the result of LBP operator will have
different value. Maenpaa extend the operator that the
result of LBP is the minimum of rotated LBP operator [5]. Another extension to the operator is the definition of
so-called uniform patterns. A local binary pattern is
called uniform if it contains at most two bitwise
transitions from 0 to 1or vice versa when the bit pattern is
considered circular [6]. For example, the patterns
Journal of Image and Graphics, Vol. 4, No. 1, June 2016
©2016 Journal of Image and Graphics 11doi: 10.18178/joig.4.1.11-14
11000011, 00111110 and 00000000 are uniform whereas
the patterns 11001001 and 01010011 are not [7].
III. LOCAL TERNARY PATTERNS
Local Ternary Patterns are new 3-valued texture
operator that can be considered as an extension to LBP.
The LTP will define a threshold a threshold say t and any
pixel value within the interval of –t and +t, thus assigns
the value 0 to that pixel [8], while the user assigns the
value 1 to that pixel if it is above this threshold and a
value -1 if it is below it when compared to the central
pixel value. The S(u) is replaced with function (2):
1, ( ) t
( , , ) 0, | |
1, ( )<-t
i c
i c i c
i c
P P
S P P t P P t
P P
(2)
To get rid of the negative values, the LTP values are
divided into two LBP channels, the upper LTP (LTPU)
and the lower LTP (LTPL). The LTPU is obtained by
replacing the negative values in the original LTP by zeros.
The LTPL is obtained in two steps, first, we replaced all
the value of 1 in the original LTP to be zeros then we
changed the negative values to be 1. See Fig. 2 for an
illustration for procedure of the LTP operator.
Figure 2. Procedure of the LTP operator.
IV. WEIGHTED LTP
About the weight of facial sub regions, in the
references [2], the weights were selected without utilizing
an actual optimization procedure and thus there are
probably not optimal. In this paper, we applied
information entropy to compute weight [4]. The
information entropy is computed as in (3).
, ,
0
log
n
i i j i j
j
E P P
(3)
where in this case j is the gray-level value, n is 255. Pi,j
Pi,j is computed as in (4).
,
1
{ ( , ) }
{ ( , ) }
i
i j m
i
T f x y j
P
T f x y j
(4)
Numerator represents number of j-th gray value in i-th
sub regions, denominator represent number of j-th gray
value in facial image. Thus, the weight of i-th sub regions
was defined as (5).
1
i
i m
i
i
E
E
(5)
According to the theory of LTP, the LTP values of
facial sub regions are divided into LTPU and LTPL. A
LTP characteristic histogram of the whole face image is
cascaded of LTPU characteristic histogram and LTPL
characteristic histogram of all sub regions. The detailed
process of extracting LTP features is shown in Fig. 3.
(a)
(b)
Figure 3. (a) Block face image and extend LTP features (b) cascade of histogram of all sub regions.
Finally, in the histogram matching, the weighted Chi
square distance can be defined as (6).
2
2 1 2
1 2
,1 2
,i i
j
i ji i
H HH H
H H
(6)
in which H1 and H2 are the histograms to be compared,
indices i and j refer to i-th bin in histogram corresponding
to the j-th sub region and wj that calculated by (5) is the
weight for region j.
V. EXPERIMENTAL RESULTS
This proposed approach is applied to ORL and Yale
face database. The ORL face database have 400 images
Journal of Image and Graphics, Vol. 4, No. 1, June 2016
©2016 Journal of Image and Graphics 12
value in i-th sub regions. gray represents probability of j-th
that consist of ten different images of each of 40 distinct
subjects, the images were taken at different times,
varying the lighting, facial expressions (open or closed
eyes) and facial details (glasses or no glasses) [9]. The
Yale face database contains 165 grayscale images of 15
individuals. There are 11 images per subject, one per
different facial expression or configuration: center-light,
glasses, happy, left-light, wink and so on [9]. The ORL
face database is shown as Fig. 4, and the Yale face
database is shown as Fig. 5.
Figure 4. ORL face database.
Figure 5. Yale face database.
The block diagram for experiment is shown as Fig. 6.
Figure 6. Block diagram of experimental.
To assess the performance of weighted LTP algorithm,
we compare this algorithm with LBP and LTP in
reference 2. In this work, 5 images are selected as
training sample per subject, and the rest of images per
subject are test images. Threshold t is set to 5 in the LTP
operator, and there are three kind of dividing face images
such as 3*3, 4*4 and 5*5.
The experimental object is the correct rate of face
recognition [10]. The performance of the face recognition
method can be measured with recognition rate, defined as
follows:
recognition rate 100%number of correct recognition
number of face
The experimental results are shown in Table I and
Table II.
TABLE I. EXPERIMENTAL RESULT IN ORL (%)
Method Block Size
3*3 4*4 5*5
LBP 86.00 88.50 93.00
LTP in [2] 88.50 92.00 96.50
Weighted LTP 89.00 93.50 97.00
TABLE II. EXPERIMENTAL RESULT IN YALE (%)
Method Block Size
3*3 4*4 5*5
LBP 87.78 91.11 93.33
LTP in [2] 90.00 93.33 96.67
Weighted LTP 91.11 96.67 97.78
We learn from Table I and Table II that the correct rate
of weighted LTP is better than LBP and LTP in reference
2. The experimental result show that the correct rate of
weighted LTP increase about 3%~4% in ORL face
database and increase about 3%~5% in Yale face
database. And the correct rate of block size of 5*5 more
8% than block size of 3*3 and more 3.5% than block size
of 4*4 in ORL face database, while more 6.67% than
block size of 3*3 and more 1.11% than 4*4 in Yale face
database.
At the same time, we take a experiment that number of
training sample produce an effect on the correct rate of
weighted LTP. Where the threshold of weighted LTP is
defined as 5 and block size is 5*5, and the number
training is 2-9 of ORL face database, 2-10 of Yale face
database, the results are described by Fig. 7 and Fig. 8.
With the increase in the number of training, the correct
recognition rate of weighted LTP is gradually increased
in ORL and Yale face database. When training number is
increased from 2 to 5, the recognition rate increased
quickly. While the recognition rate increased slowly with
the training number increased from 5 to 9.
Figure 7. Experiment of ORL face database.
Journal of Image and Graphics, Vol. 4, No. 1, June 2016
©2016 Journal of Image and Graphics 13
Figure 8. Experiment of Yale face database
VI. CONCLUSIONS
In this paper, we make an improvement that calculates
histogram similarity combined with weight in LTP
operator. So, the method can better to extract facial
texture features and robustness to illumination. In order to
validate the effective of our proposed method, we applied
two face database for recognition. Experiment on the
ORL and Yale face database show that the weighted LTP
achieve much better performance than LBP and LTP in
reference 2.
Unfortunately, defects still exist. We need new insights
into the role of method of face recognition played in
dealing with difficult lighting conditions and difficult
noise in face image. For future work, more discriminative
features will be researched to improve the robustness and
accuracy of face recognition.
REFERENCES
[1] T. Ahonen, A. Hadid, and M. Pietikainen, “Face description with local binary pattern: Application to face recognition,” IEEE
Transactions on Patterns Analysis and Machine Intelligence, vol. 28, pp. 2037-2041, Dec. 2006.
[2] X. Tan and B. Triggs, “Enhanced local texture feature sets for face
recognition under difficult lighting conditions,” IEEE Transactions on Image Processing, vol. 19, pp. 1635-1650, June
2010. [3] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution gray
scale and rotation invariant texture classification with local binary
patterns,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, pp. 971-987, July 2002.
[4] C. A. Perez, L. A. Cament, and L. E. Castillo, “Local matching Gabor entropy weighted face recognition,” in Proc. IEEE
International Conference, March 2011, pp. 179-184.
[5] X. H. Han, G. Xu, and Y. W. Chen. “Robust local ternary patterns for texture categorization,” in Proc. 6th International Conference
on Biomedical Engineering and Informatics, June 2013, pp, 846-850.
[6] W. H. Liao, “Region description using extended local ternary
pattern,” in Proc. International Conference on Pattern Recognition, Istanbul, Turkey, 2010, pp, 1003-1006.
[7] J. Ren, X, Jiang, and J. Yuan, “Relaxed local ternary pattern for face recognition,” in Proc. International Conference on
Information Processing, 2013, pp. 3680-3684.
[8] W. H. Liao and T. J. Young, “Texture classification using uniform extended local ternary patterns,” in Proc. International Symposium
on Multimedia, 2010, pp, 191-195. [9] ORL and Yale face databases. [Online]. Available:
http://www.face-rec.org/databases/
[10] J. Meng, Y. Gao, X. Wang, T. Lin, and J. Zhang, “Face recognition based on local binary patterns with threshold,” in Proc.
IEEE International Conference on Granular Computing, San Jose,
CA, 2010, pp. 352-356.
Haifeng Zhang has received Master’s Degree from Zhejiang University in Thermal Physics. Now, he is a master instructor and associate
professor in Hangzhou Dianzi University. His research interests focus on signal processing, RF, and embedded system.
Journal of Image and Graphics, Vol. 4, No. 1, June 2016
©2016 Journal of Image and Graphics 14
Shenjie Xu is a studying master of Hangzhou Dianzi University, Hangzhou, China. His research interests include embedded system and
face recognition.